<p>In the textile industry, the ability to effectively detect multiple types of fabric anomalies remains a core challenge. Traditional student–teacher networks are constrained by their fixed receptive fields, making it difficult to effectively detect large-area texture anomalies, especially color anomalies. To overcome this bottleneck, this paper proposes a dual-student–teacher network architecture, with the patch description network and masked autoencoder selected as the backbones of the two student networks. Our research demonstrates that the masking strategy of the masked autoencoder enables the model to fully utilize contextual information from the visible region, allowing it to exhibit excellent complementarity with the patch description network in fabric anomaly detection tasks. Secondly, this paper introduces a focus pixel fusion strategy that adaptively adjusts the fusion weights and fusion ranges of feature maps based on the anomaly types, achieving refined information integration between different student networks. Finally, we propose the PNFE image preprocessing method, which mitigates the impact of overfitting on the network by adding Perlin noise to fabrics to generate random shadows and applying random fisheye effects, improving training performance without increasing the number of training images. Compared with other methods from recent years, the experimental results of AFF-DSTnet validate its superiority in relevant metrics.</p>